h mSSKrSSKrSSKrSSKrSSKrSSKJr SSKJr SSKJ r SSK J r J r J r Jr SSKrSSKrSSKJrJr SSKJrJrJr SS /r"S S \R05r"S S \5rSrS\S\4SjrS\\4SjrS\ \ \!\ 4S\ \!\ 44Sjr"S!S\RFS\ \ \!\ 4S\$S\ \ \!\ 4S\%\\!\\!44 Sjjr&"SS\RN5r(S"S\RFS\ \\\!\ \!\!44S\ \\\!\ \!\!44S\ \\\!\ \!\!44S\ \ \!\ 4S\$S\ \ \!\ 4S\RN4S jjr)g)#N) OrderedDict)deepcopy)chain)AnyCallableOptionalUnion)fxnn)_CodeOnlyModule _copy_attr_USER_PRESERVED_ATTRIBUTES_KEYcreate_feature_extractorget_graph_node_namesc\^\rSrSrSrU4SjrS\RS\S\ 4U4Sjjr Sr U=r $) LeafModuleAwareTracerz An fx.Tracer that allows the user to specify a set of leaf modules, i.e. modules that are not to be traced through. The resulting graph ends up having single nodes referencing calls to the leaf modules' forward methods. cn>0UlSU;aURS5nX0l[TU] "U0UD6 g)N leaf_modules)rpopsuper__init__)selfargskwargsr __class__s ]/var/www/auris/envauris/lib/python3.13/site-packages/torchvision/models/feature_extraction.pyrLeafModuleAwareTracer.__init__s9 V #!::n5L ,  $)&)mmodule_qualnamereturncl>[U[UR55(ag[TU]X5$)NT) isinstancetuplerris_leaf_module)rr r!rs rr&$LeafModuleAwareTracer.is_leaf_module"s. at001 2 2w%a99r)r) __name__ __module__ __qualname____firstlineno____doc__rr Modulestrboolr&__static_attributes__ __classcell__rs@rrrs0 *: :C:D::rrc^\rSrSrSrU4SjrS\RRS\ 4Sjr SS\ S\ RRS \ RR 4U4S jjjrS \ S \ RR$S \ 4S jrSrU=r$)NodePathTracer(a, NodePathTracer is an FX tracer that, for each operation, also records the name of the Node from which the operation originated. A node name here is a `.` separated path walking the hierarchy from top level module down to leaf operation or leaf module. The name of the top level module is not included as part of the node name. For example, if we trace a module whose forward method applies a ReLU module, the name for that node will simply be 'relu'. Some notes on the specifics: - Nodes are recorded to `self.node_to_qualname` which is a dictionary mapping a given Node object to its node name. - Nodes are recorded in the order which they are executed during tracing. - When a duplicate node name is encountered, a suffix of the form _{int} is added. The counter starts from 1. cR>[TU]"U0UD6 SUl[5Ulg)N)rrcurrent_module_qualnamernode_to_qualname)rrrrs rrNodePathTracer.__init__;s) $)&)')$ !, rr forwardcURnURU5nX`lURX5(dU"U0UD6nUXPl$URSXcU5XPl$!XPlf=f)a_ Override of `fx.Tracer.call_module` This override: 1) Stores away the qualified name of the caller for restoration later 2) Adds the qualified name of the caller to `current_module_qualname` for retrieval by `create_proxy` 3) Once a leaf module is reached, calls `create_proxy` 4) Restores the caller's qualified name into current_module_qualname call_module)r8path_of_moduler& create_proxy)rr r;rr old_qualnamer!outs rr=NodePathTracer.call_moduleEsv33  8"11!4O+: (&&q::t.v.,8 ($$]O6R+7 (< (s6A$ A$$A,kindtargetr"c>[T U]XX4XV5nURURUR5UR UR'U$)z Override of `Tracer.create_proxy`. This override intercepts the recording of every operation and stores away the current traced module's qualified name in `node_to_qualname` )rr?_get_node_qualnamer8noder9) rrCrDrrname type_expr_proxyrs rr?NodePathTracer.create_proxyZsK$T4Q,0,C,CDD`D`bgblbl,mejj) rr!rGcUnURS:wa"[U5S:aUS- nU[U5- n[R"SU5bUR SS5Sn[ URR55H_n[R"US3U5cM URUS5n[U5(a[USS5S-nOSnUSU3- n U$ U$) Nr=r.z .+_[0-9]+$rJz (_[0-9]+)?$r7) oplenr.rematchrsplitreversedr9valuesreplaceint)rr!rG node_qualnameexisting_qualnamepostfix next_indexs rrF!NodePathTracer._get_node_qualnamefs' 77m #=!A%$ SY &M 88M= 1 =)00a8;M"*$*?*?*F*F*H!I xxM?+68IJV+33M2Fw<^# UHm[U4SjT55v M g7f)c3,># UH oT:Hv M g7fN).0x_itemy_items r '_is_subseq...s9&V#&sN)any)rgriiter_xs @rrj_is_subseq..sJfs9&999s"&)iterall)xyrms @r _is_subseqrss!WF JJ JJr train_tracer eval_tracerc[URR55n[URR55n[U5[U5:Xa![ S[ X#555(agSn[ X#5(aSnO[ X25(aSnOSn[R"XT-5 g)zx Utility function for warning the user if there are differences between the train graph nodes and the eval graph nodes. c3.# UH upX:Hv M g7frerf)rgtes rrj*_warn_graph_differences..s2cFbda16FbsNztWhen choosing nodes for feature extraction, you may need to specify output nodes for train and eval mode separately.znNOTE: The nodes obtained by tracing the model in eval mode are a subsequence of those obtained in train mode. znNOTE: The nodes obtained by tracing the model in train mode are a subsequence of those obtained in eval mode. zdThe nodes obtained by tracing the model in train mode are different to those obtained in eval mode. ) listr9rVrQrpziprswarningswarn)rtru train_nodes eval_nodessuggestion_msgmsgs r_warn_graph_differencesrs |44;;=>Kk2299;` in the `documentation `_. Args: model (nn.Module): model for which we'd like to print node names tracer_kwargs (dict, optional): a dictionary of keyword arguments for ``NodePathTracer`` (they are eventually passed onto `torch.fx.Tracer `_). By default, it will be set to wrap and make leaf nodes all torchvision ops: {"autowrap_modules": (math, torchvision.ops,),"leaf_modules": _get_leaf_modules_for_ops(),} WARNING: In case the user provides tracer_kwargs, above default arguments will be appended to the user provided dictionary. suppress_diff_warning (bool, optional): whether to suppress a warning when there are discrepancies between the train and eval version of the graph. Defaults to False. concrete_args (Optional[Dict[str, any]]): Concrete arguments that should not be treated as Proxies. According to the `Pytorch docs `_, this parameter's API may not be guaranteed. Returns: tuple(list, list): a list of node names from tracing the model in train mode, and another from tracing the model in eval mode. Examples:: >>> model = torchvision.models.resnet18() >>> train_nodes, eval_nodes = get_graph_node_names(model) rrf) rtrainingr4tracetrainevalr{r9rVr) rrrr is_trainingrtrurrs rrrsd/}=M..K!2M2Lu{{}MB 1=1Kejjl-@|44;;=>Kk2299;[[RU]5 X@RlX lX0l[[UR5[UR55HknURS;dM[UR[5(d![S[!UR535e[#XUR5 Mm UR%5 X lURR(UR R(:wa:[SURR(SUR R(S35eSUlUR&R((aASUR&R(R*;aUR&R(Ulggg)z Args: root (nn.Module): module from which the copied module hierarchy is built train_graph (fx.Graph): the graph that should be used in train mode eval_graph (fx.Graph): the graph that should be used in eval mode )get_attrr=z-node.target should be of type str instead of zGTrain mode and eval mode should use the same tracer class. Instead got z for eval vs z for trainNz)rr GraphModulerrr(rrrronodesrPr$rDr. TypeErrortyper rgraph _tracer_clsr*)rrrrrrGrs rrDualGraphModule.__init__s{ bnnd,.",&$${0014 8H8H3IJDww55!$++s33#&STXY]YdYdTeSf$ghh4t{{3 K   ?? & &$*:*:*F*F FYZ^ZiZiZuZuYvwDEIEUEUEaEaDbblm   :: ! !j 8N8N8[8[&[#zz55D '\ !rc>U(a#UR(dURUlO)U(d"UR(aURUl[TU]US9$)z Swap out the graph depending on the selected training mode. NOTE this should be safe when calling model.eval() because that just calls this with mode == False. )mode)rrrrrr)rrrs rrDualGraphModule.train@s@  ))DJ$--DJw}$}''rc"[R$re)rr)rs r_deepcopy_initDualGraphModule._deepcopy_initNs'''rc [U5R[U55nX![U5'[[R "UR U55nUR5"X#UR SUR S5 /SQnUHBnXPR ;dM[X%[R "UR UU55 MD [R "[US05U5Ul [UR;a6UR[R5Hupe[X&U5 M U$)Nrr)_state_dict_hooks_load_state_dict_pre_hooks_load_state_dict_post_hooks _replace_hook_create_node_hooks_erase_node_hooksmeta) r__new__idr copyr__dict__rsetattrgetattrrritems)rmemoresfake_modextra_preserved_attrsattr attr_names r __deepcopy__DualGraphModule.__deepcopy__Rs4j  d,RX"4==#EF cX->->}-MxO`O`amOno! *D}}$4==t1Dd#KL*==vr!:DA )SXX 5#&88,J#K#Q#Q#S -$T r)rrrr)r)T)r(r)r*r+r,r^r r-r Graphr.rrrrr0r1r2s@rrr s^er)6HHOO)624(()6HJ)6^a)6)6V ((rr return_nodestrain_return_nodeseval_return_nodesc [U5nURn[SXU455(a [S5eUSLUSL- (a [S5eUSLUSL- (d [S5eS[[ [ 44SjnUcU"U5n[ U5n[ U5nOU"U5nU"U5n0n 0n X#S.n SGHHn U S :XaUR5 OU S :XaUR5 [S0UD6n U RXS 9n[U[R5(aURRO URn[ R""U R$X5n['U R(R+55n[-[/U55[-U5:wa [S 5eXR15HNn[3UVs/sH n[4R6"S US3U5SLPM" sn5(aMA[SUS35e /n[9UR:R<5H&nUR>S:XdMURAU5 M( U(d [S5eUHnUR:RCU5 M UR:R<Vs/sHnUPM nn[E5n[9U5HnU R(RGU5nUcM#XH^nURIS5nSRKURMS5SUS-5U:XdM@UUXU'XROU5 M M [E[9['URQ5555nUR:RSUS5 UR:RUU5 SSS5 UR:RW5 URY5 XU 'XU 'GMK U(d[[U S U S 5 []X S U S WS9nURU5 URU5 U$s snfs snf!,(df  N=f)a Creates a new graph module that returns intermediate nodes from a given model as dictionary with user specified keys as strings, and the requested outputs as values. This is achieved by re-writing the computation graph of the model via FX to return the desired nodes as outputs. All unused nodes are removed, together with their corresponding parameters. Desired output nodes must be specified as a ``.`` separated path walking the module hierarchy from top level module down to leaf operation or leaf module. For more details on the node naming conventions used here, please see the :ref:`relevant subheading ` in the `documentation `_. Not all models will be FX traceable, although with some massaging they can be made to cooperate. Here's a (not exhaustive) list of tips: - If you don't need to trace through a particular, problematic sub-module, turn it into a "leaf module" by passing a list of ``leaf_modules`` as one of the ``tracer_kwargs`` (see example below). It will not be traced through, but rather, the resulting graph will hold a reference to that module's forward method. - Likewise, you may turn functions into leaf functions by passing a list of ``autowrap_functions`` as one of the ``tracer_kwargs`` (see example below). - Some inbuilt Python functions can be problematic. For instance, ``int`` will raise an error during tracing. You may wrap them in your own function and then pass that in ``autowrap_functions`` as one of the ``tracer_kwargs``. For further information on FX see the `torch.fx documentation `_. Args: model (nn.Module): model on which we will extract the features return_nodes (list or dict, optional): either a ``List`` or a ``Dict`` containing the names (or partial names - see note above) of the nodes for which the activations will be returned. If it is a ``Dict``, the keys are the node names, and the values are the user-specified keys for the graph module's returned dictionary. If it is a ``List``, it is treated as a ``Dict`` mapping node specification strings directly to output names. In the case that ``train_return_nodes`` and ``eval_return_nodes`` are specified, this should not be specified. train_return_nodes (list or dict, optional): similar to ``return_nodes``. This can be used if the return nodes for train mode are different than those from eval mode. If this is specified, ``eval_return_nodes`` must also be specified, and ``return_nodes`` should not be specified. eval_return_nodes (list or dict, optional): similar to ``return_nodes``. This can be used if the return nodes for train mode are different than those from eval mode. If this is specified, ``train_return_nodes`` must also be specified, and `return_nodes` should not be specified. tracer_kwargs (dict, optional): a dictionary of keyword arguments for ``NodePathTracer`` (which passes them onto it's parent class `torch.fx.Tracer `_). By default, it will be set to wrap and make leaf nodes all torchvision ops: {"autowrap_modules": (math, torchvision.ops,),"leaf_modules": _get_leaf_modules_for_ops(),} WARNING: In case the user provides tracer_kwargs, above default arguments will be appended to the user provided dictionary. suppress_diff_warning (bool, optional): whether to suppress a warning when there are discrepancies between the train and eval version of the graph. Defaults to False. concrete_args (Optional[Dict[str, any]]): Concrete arguments that should not be treated as Proxies. According to the `Pytorch docs `_, this parameter's API may not be guaranteed. Examples:: >>> # Feature extraction with resnet >>> model = torchvision.models.resnet18() >>> # extract layer1 and layer3, giving as names `feat1` and feat2` >>> model = create_feature_extractor( >>> model, {'layer1': 'feat1', 'layer3': 'feat2'}) >>> out = model(torch.rand(1, 3, 224, 224)) >>> print([(k, v.shape) for k, v in out.items()]) >>> [('feat1', torch.Size([1, 64, 56, 56])), >>> ('feat2', torch.Size([1, 256, 14, 14]))] >>> # Specifying leaf modules and leaf functions >>> def leaf_function(x): >>> # This would raise a TypeError if traced through >>> return int(x) >>> >>> class LeafModule(torch.nn.Module): >>> def forward(self, x): >>> # This would raise a TypeError if traced through >>> int(x.shape[0]) >>> return torch.nn.functional.relu(x + 4) >>> >>> class MyModule(torch.nn.Module): >>> def __init__(self): >>> super().__init__() >>> self.conv = torch.nn.Conv2d(3, 1, 3) >>> self.leaf_module = LeafModule() >>> >>> def forward(self, x): >>> leaf_function(x.shape[0]) >>> x = self.conv(x) >>> return self.leaf_module(x) >>> >>> model = create_feature_extractor( >>> MyModule(), return_nodes=['leaf_module'], >>> tracer_kwargs={'leaf_modules': [LeafModule], >>> 'autowrap_functions': [leaf_function]}) c3(# UHoSLv M g7frerf)rgargs rrj+create_feature_extractor..s X"W3$;"WszcEither `return_nodes` or `train_return_nodes` and `eval_return_nodes` together, should be specifiedNzcIf any of `train_return_nodes` and `eval_return_nodes` are specified, then both should be specifiedz\If `train_return_nodes` and `eval_return_nodes` are specified, then both should be specifiedr"c[U[5(a&UVs0sHn[U5[U5_M sn$UR5VVs0sHup#[U5[U5_M snn$s snfs snnfre)r$r{r.r)nikvs r to_strdict,create_feature_extractor..to_strdictsa a  ,-.AqCFCFNA. .+,7795941AA955/5s A4!A9)rrrrrzYThere are duplicate nodes! Please raise an issue https://github.com/pytorch/vision/issues^z(\.|$)znode: 'z' is not present in model. Hint: use `get_graph_node_names` to make sure the `return_nodes` you specified are present. It may even be that you need to specify `train_return_nodes` and `eval_return_nodes` separately.outputz1No output nodes found in graph_module.graph.nodesrNrO)rrf)/rrrp ValueErrordictr.rrrr4rr$r r-rr(r rrr{r9rVrQrkeysrlrRrSrUrrrPr erase_nodergetcountjoinsplitrrinserting_afterreliminate_dead_code recompilerr)rrrrrrrrrtracersgraphsmode_return_nodesrtracerrrH graph_moduleavailable_nodesqueryrorig_output_nodesr output_nodesr!depths rrrqs6j/}=M..K XVZ[L KK {# ^*6&; :s'Q ; Q"4Q'' Q5 )NFN)NNNNFN)*rrrrRr} collectionsrr itertoolsrtypingrrrr r^rr r torch.fx.graph_moduler r r__all__Tracerrr4rsrr{rrrr.rr-r/r%rrrrrfrrrs  #11 ]] &'= >:BII:(`*`FK(.(~(>4: 8DcN3K PTUXZ]U]P^ &/3"'.2 ># 99>#DcN+># >#DcN+ >#  49d3i  >#BabnnaL@DEIDH.2"'.2n 99n5cDcN!:;<n!tCy$sCx.'@!ABn d3ic3h&? @A n DcN+ n  nDcN+n^^nr